Run multiple AI coding agents in parallel terminal tiles

kolu is a terminal app built for running many things at once: real xterm.js terminal tiles arranged freely on an infinite 2D canvas, with a dock that keeps track of every one so none gets lost. It works with any agent CLI you run in a shell - Claude, Codex, OpenCode, Grok, Pi, or anything else - rather than wrapping one model behind its own chat UI.

Use it when you're running several terminal-based coding agents or shell sessions in parallel and need to actually see and switch between them - the canvas/tile/dock model is built for arranging many real terminals and finding any one instantly, and an agent-attention view shows which agents are working, finished, or waiting for input. It's explicitly agent-agnostic: there's no agent registry or adapter to configure, and kolu instead auto-detects the repositories, agents, and sessions you already use rather than asking you to set them up twice. Sessions are durable - shells and agents stay alive across tab closes, server restarts, and redeploys - and kolu can reach across machines: access it remotely from another device, add another machine onto the same canvas, and forward a port so a dev server running on any connected machine opens directly in your browser. The typical setup is a Mac window onto a headless Linux box, though it can also run as a home-manager service on Linux or macOS directly.

Input is whatever shell commands or agent CLI invocations you run inside a terminal tile; output is that terminal's live output, rendered in the browser workspace alongside every other tile on the canvas. The workspace itself - tile layout, session state, and live PTYs - persists across browser tab closes and server restarts rather than resetting.

nix --refresh run github:juspay/kolu -- web

Requires Nix with flakes enabled; running that command starts kolu and serves the workspace at http://127.0.0.1:7681. If the first run starts compiling instead of downloading, kolu's binary cache needs to be enabled first. Under the hood, the browser workspace, web shell, workspace state, and live PTYs run as separate processes that can fail and restart independently rather than taking the whole session down together.

Developers running multiple terminal-based AI coding agents or shell sessions side by side who want one persistent, remotely reachable workspace to arrange and monitor all of them, without being locked into a single vendor's agent or chat interface.